Stacking ensemble for multilingual aspect-based sentiment analysis of tourism reviews

Abstract

Multilingual tourism reviews often express opinions about multiple destination attributes within a single text, making aspect-level sentiment classification challenging. This study compares classical machine learning and ensemble methods for multilabel aspect-based sentiment analysis of amenities, attractions, and accessibility. The dataset comprises 13,085 Indonesian- and English-language Google Reviews labeled through a rule-based procedure. Reviews were represented using term frequency-inverse document frequency features and classified using logistic regression, multinomial naive Bayes, random forest, soft voting, adaptive boosting (AdaBoost), and stacking under a one-vs-rest framework. The stacking classifier combined logistic regression and random forest as base learners with logistic regression as the meta-learner. It achieved the best performance among the evaluated models, with a micro-F1 of 0.8626, macro-F1 of 0.7285, Hamming loss of 0.0633, and subset accuracy of 0.6825. These findings indicate that stacking effectively integrates complementary decision patterns and provides a reproducible framework for multilingual multilabel analysis of tourism reviews.